Adversarial Research

Peer-review-by-agent. Inspired by Mark Kashef's second dynamic-workflow use case (YT 9_ExDZFlaNc): a single deep-research agent claims "checked 300 sources" but never plays devil's advocate against itself. This spins up a separate verification swarm whose job is to break each claim.

When to run

You have a body of contested claims (a new feature's capabilities, a vendor's marketing, a hyped technique) and want signal separated from hype. Token-heavy — Mark's run was ~200 agents / 20 min. Scope the depth.

Relationship to your existing skills: deep-recon finds existing solutions (prior-art search); deep-research synthesizes a cited report. This one is specifically the adversarial verification layer — use it on top of, or instead of, those when claims are disputed.

The recipe

1. Gather claims

Pull raw claims from the relevant surface:

• X/Twitter — Apify X scraper actor (search apify MCP: search-actors → tweet scraper). You already use Apify for YouTube analytics. Alternatively the native X API toolkit (~/Developer/personal/x-tools/).
• Web/GitHub — mcp__exa__web_search_advanced_exa, mcp__exa__crawling_exa, mcp__ref__*.

Extract atomic claims (one assertion each) with their source URL. Write to a claims file.

2. Adversarial verification (Workflow tool, explicit opt-in)

Pipeline each claim through independent skeptic agents:

pipeline(claims,
  claim => agent(`Try to REFUTE this claim. Default to refuted=true if you cannot find
                  primary evidence. Claim: ${claim.text}  Source: ${claim.url}`,
                 {schema: VERDICT}),
  ...
)

For high-stakes claims use a perspective-diverse panel (3 verifiers, distinct lenses: primary-source / reproducibility / counter-example) and require majority to survive.

3. Label + report

Each claim → survives (independently verified) / cut (refuted) / unconfirmed (no primary evidence either way), with the link that settled it. Write an HTML report grouping by verdict so disputed claims are scannable with "a second set of eyes."

Guardrails

• Verifiers must cite the source that confirmed/refuted — no unsourced verdicts.
• Bias toward unconfirmed over survives when evidence is thin. Hype masquerades as confirmed.
• Route web fetches through the sandboxed research proxy where appropriate (research-proxy / link-preview-subagent) to avoid indirect prompt injection from scraped content.

Workflow reliability (learned 2026-06-02)

CRITICAL: the verify agents do web research and end in prose, so forcing a schema on them failed catastrophically (~190/239 dropped to null on the first run). Two safe fixes: (1) have each verifier return a one-line verdict in plain text like VERDICT: survives | url | reason and parse with a regex/jq pass, or (2) end the prompt with "Your FINAL action MUST be the StructuredOutput call; do not end your turn with prose." Also cap extraction — 80 tweets produced 239 fragmentary "claims"; instruct extractors to emit only distinct, substantive, testable claims (target ≤ ~5/tweet). See [[feedback_workflow_structured_output_failure]].